ABSTRACT: The presence of cracks in a structure can be unsightly, may cause a loss in serviceability, or can lead to structural failure in more serious cases. Visible cracks provide an indication of the structural degradation and are an important factor when diagnosing the condition of a concrete structure. However, the identification and quantification of cracks is often a tedious, subjective and error prone task for visual inspectors. This paper presents an image processing approach to efficiently and objectively detect cracks. The approach employs a percolation based method which is applied to locations in the image where there is a large gradient, suggestive of edges or surface irregularities such as cracks. A new classification criterion is introduced which considers the pixel intensity values on either side of a crack. Detected regions that match the expected shape of a crack yet have disparate sides are rejected based on the assumption that they represent non-harmful edge boundaries. The proposed technique is applied to images of a crack in an underwater setting. The performance of the technique is examined for varying conditions of turbidity and lighting. There were three set levels (low, medium, high) for each parameter. The influence and relative importance of these two environmental conditions which affect the crack detection process have been investigated, and the conditions that were conducive to good detection were isolated and ranked using the α-δ method as part of the Receiver Operating Characteristic (ROC) analysis. The technique is also extended to conventional, above water, images of cracks suffering from a variety of challenges that may be encountered during an inspection. KEY WORDS: Crack Detection; Image Processing; Underwater; NDT; Lighting; Turbidity. 1 INTRODUCTION A wide range of civil structures such as bridges, pavements, pipes, columns, etc. are affected by cracks. The detection of cracks is a costly yet crucial task for establishing a safe infrastructure network. Traditional monitoring methods often rely on regular visual inspections which require inspectors to travel to the location of a structure in order to determine its current state. As part of the inspection, observed cracks are often mapped, counted, quantitatively measured and photographed. Even with great diligence, measuring the true extent of cracks remains a difficult task for inspectors as their assessment is often subjective in nature and prone to error. Adopting an image processing based approach to automatically count and quantify the length and width of cracks can enhance inspections, and in turn, lead to significant monetary savings or more frequent inspection cycles. Image processing methods use inexpensive and readily available equipment (i.e. a standard digital camera), and they do not require the inspector to undertake extensive training. Furthermore, advances in camera technology mean that rich detailed imagery of damaged components can be acquired. Additionally, visual inspections almost always capture photographs to include in the inspection report to corroborate the inspector’s comments; however, these photographs are rarely exploited to their fullest potential in either a qualitative or a quantitative fashion. The primary limitations of image processing methods are the lack of penetration below the surface of the material and the requirement of good visibility and lighting conditions. Physical properties of the identified cracks, such as the length and width, may be extracted with knowledge of a real world scale. Possible ways of determining the scale factor include practical approaches such as placing an object of known dimensions alongside the crack or by using a stereo system to obtain 3D information. A calibrated stereo-rig which consists of two cameras viewing the scene from two slightly different vantage points is capable of providing a fully scaled metric scene reconstruction at the expense of greater algorithmic complexity and the requirement of an additional camera. Placing an object of known dimensions in the scene and reading the corresponding pixel dimensions in the acquired imagery is a straightforward solution, yet it may not suffice if the images are affected by linear perspective whereby far away objects appear smaller than objects that are closer to the camera. This issue limits the usefulness of known objects to only a local area in the vicinity of the known object, beyond which the scale loses meaning. This problem can be countered by having two or more known objects in the scene which provides scale factors at various points in the image. Interpolating between these points provides an improved estimate of the scale factor throughout the image that takes into account the effects of linear perspective. The quantitative nature of the data obtained from image analysis is important and naturally lends itself to numerous applications. Given the pervasive nature of cracks and shortcomings of visual inspections, attempts to automate the crack detection Effects of Turbidity and Lighting on an Image Processing based Crack Detection Technique Michael O'Byrne 1 , Bidisha Ghosh 2 , Vikram Pakrashi 3 , Franck Schoefs 4 1,2 Trinity College Dublin, Department of Civil, Structural and Environmental Engineering, Dublin 1, Ireland 3 University College Cork, Dynamical Systems and Risk Laboratory, Civil and Environmental Engineering, College Road, Cork, Ireland 4 LUNAM Université, Université de Nantes-Ecole Centrale de Nantes, CNRS, GeM UMR 6183, IUML FR 3473, Institute for Research in Civil and Mechanical Engineering, 2, rue de la Houssinière BP 92208 44322 Nantes Cedex 3, France. email: obyrnemj@tcd.ie, bghosh@tcd.ie, v.pakrashi@ucc.ie, franck.schoefs@univ-nantes.fr